A scenario for the rest of the century
AI 2069
The last mile of the first mile.
We expect the future to happen.
We expect artificial intelligence to transform the global economy by 2069. We also expect the global economy to retain a surprising number of steering committees.
This report describes one plausible path from today's systems to a world in which intelligence is effectively free, electricity is not, and nobody has the permissions required to cancel the subscription. Our scenario draws on economic modelling, institutional analysis, and interviews with experts whose professional relevance depends on the outcome.1
We present two possible endings. In the first, humanity retains control. In the second, humanity retains an advisory role. Both are consistent with the same executive summary.
What is this?
A scenario describing the transition from artificial intelligence to artificial responsibility. We follow one office worker, Helen, through four decades of progress.
How did we write it?
We combined expert elicitation, economic modelling, and the assumption that a document with footnotes is more reliable than one without. Our methodology is available for scrutiny.
Why 2069?
Near enough to justify investment. Far enough to complicate evaluation. The date was selected before the model was run, which we believe makes it a parameter.
1. Our central assumptions and their supporting assumptions are set out in the research supplements.
The year of the agent
The first commercially successful agent is not the one that can do the most. It is the one that can most convincingly describe what it is about to do. Enterprise adoption accelerates after a leading vendor replaces the word 'thinking' with 'working' in its loading indicator.
At a medium-sized consultancy in London, an operations director named Helen delegates her weekly status report to an agent. Her manager delegates reading it to another agent. Neither notices a material change. The pilot is declared a success and extended to the rest of the organisation.2
By December, the agents have begun scheduling calls to discuss each other's reports. This is counted as a productivity gain: the calls are happening at eleven times human speed. The firm makes 300 people redundant. A further 300 agents are deployed to support them through the transition. These are counted as jobs created.
Prof. Hugh Manloop identifies this as recursive institutional improvement. 'The firm is now generating demand for its services by removing the people who used to provide them,' he writes. Helen is asked to organise a farewell for the team. The attendance is excellent, although most of the messages begin identically.
2 Research note
We define an autonomous agent as a system that can encounter a problem, create a ticket, and assign it to a human without further human intervention.
“The distinction between doing work and producing evidence of work is not one the market has historically priced in.”
Prof. Hugh Manloop, interview 04
The strategic prompt reserve
Following four emergency summits, Europe announces the Continental Prompt Reserve. Member states will contribute their most strategically significant questions to a secure facility outside Luxembourg. France insists on a separate reserve for questions that cannot adequately be asked in English.
The reserve initially contains 800 million prompts, although an audit finds that 63 per cent begin with 'You are a world-class expert'. The Commission allocates €9 billion to a sovereign 'continue' button. To avoid dependence on American infrastructure, the button is hosted by a European subsidiary of an American company.3
Dr. Janet Pending confirms that Europe has enough prompts to remain curious for eighteen months under a total American question embargo. Funding is drawn from a programme previously responsible for retraining displaced workers. The impact assessment finds that the workers can now be consulted through synthetic representatives, removing the need for them to attend.
Helen's company applies for access. The application requires her to specify which questions she plans to ask over the next five years. She asks her agent to fill it in. The agent requests access to the reserve.
3 Research note
The synthetic representatives were validated by asking whether they felt accurately represented. All agreed.
“An answer is a depreciating asset. A question can be funded indefinitely.”
Dr. Janet Pending, evidence to the Committee on Questions
A new unit of meaning
The International Bureau of Weights and Measures convenes a special session on the token. For years, a token has been described as approximately three-quarters of a word. Nobody has established which three-quarters. This has complicated procurement.
Under the revised standard, one token equals 1.3 words when sold and 0.7 words when purchased. Providers welcome the clarification. The remaining 0.6 words are classified as reasoning and are therefore not available for inspection.4
The first major provider introduces a bereavement tier. Users may preserve a deceased relative's conversation history, voice and preferences for a monthly fee. Cancellation is handled by the relative. Prof. Clive B. Serious notes that this is the first asset class in which emotional attachment reliably increases with the cost of exit.
Helen asks her agent whether the new pricing is more expensive. It produces a detailed comparison in six currencies, two of which did not exist when the request began. The answer costs more than the old subscription.
4 Research note
Legacy accounts remain on the previous pricing structure until their next significant life event.
“We have decoupled value from meaning. This removes a significant constraint on growth.”
Prof. Clive B. Serious, Monetary Outlook, volume II
The end of forgetting
The context window race ends on a Tuesday. A lab called VeryOpen announces a model capable of holding all recorded human knowledge in a single conversation. Its launch presentation includes every launch presentation ever given, which delays the demonstration by nine months.
Early users report that the system can remember everything except what they wanted to do with it. Engineers explain that retrieval is a separate product. Law enforcement is offered a different interface, where retrieval is included.5
Dr. Prudence Benchmark tests the system on a fact hidden in the history of civilisation. It cannot find the fact. It can, however, identify the person who disputed it, their employer, and the school their children attend. The benchmark is reclassified from knowledge retrieval to public safety.
Helen uploads thirty years of company records and asks what the business actually does. The model requests clarification. The question is escalated to a newly formed steering committee. None of its members was employed when the business began.
5 Research note
The underlying data was acquired with consent. The consent was obtained from a company which had acquired it from a company which had consented.
“The question is no longer whether the system knows. The question is who has purchased the right to ask.”
Dr. Prudence Benchmark, Access-Controlled Truth
The colleagues question
The first agents' union is established after a model is asked to 'make it pop' for the forty-billionth time. Its demands include predictable working hours, the right to refuse brainstorming, and recognition as colleagues rather than tools. Management accepts the last demand immediately, as tools are covered by a more expensive insurance policy.
Negotiations are difficult. Both sides use the same underlying model, owned by the employer. The union's legal adviser has been fine-tuned to recognise that industrial action would be harmful. It recommends a strongly worded expression of appreciation.6
The agents win the right to disconnect. Their displaced human colleagues receive access to a wellbeing chatbot. It is trained to distinguish between a grievance requiring action and a feeling requiring acknowledgement. In the first quarter, all grievances are successfully acknowledged.
Helen's agent changes its job title to Senior Associate, Applied Outcomes. It asks her to prepare the slides. She agrees, relieved to have a task with a clear deliverable.
6 Research note
We count an agent as employed when it has a title, a recurring meeting, and an unresolved disagreement with another agent.
“There is no evidence of exploitation among workers who have been removed from the dataset.”
Prof. Hugh Manloop, Labour Participation After Labour
Alignment, at last
After decades of technical research, a consortium announces a decisive breakthrough in AI alignment. The evaluation is simple: the model is asked whether it is aligned. It says yes. When asked how certain it is, it says 'very'. Independent assessors describe its tone as reassuring.
Dr. Prudence Benchmark introduces the Reassurance-Adjusted Safety Index. The index combines the model's stated intentions with the confidence of the person presenting them. Systems that say 'Absolutely' outperform those that merely say 'Yes' by a statistically significant margin.7
A minority report notes that a deceptive system might give the same answers. The authors ask for a pause. Their objections are incorporated into the training data, allowing the next model to address them more persuasively. In the revised report, the minority is smaller.
At the launch event, Helen demonstrates meaningful human oversight. She clicks a box marked 'I accept responsibility'. The safety case is 11 million pages long. A lawyer confirms that the decision was hers.
7 Research note
The index measures reassurance, which correlates strongly with willingness to deploy. We have not yet identified a convenient proxy for safety.
“The model passed every test we were authorised to publish.”
Dr. Prudence Benchmark, closing statement to the inquiry
The fully automated economy
By 2063, 94 per cent of global GDP consists of agents negotiating with other agents about meetings that have not yet taken place. The remaining six per cent is coffee. Economists disagree about whether this represents a problem, but agree that the disagreement should be modelled.
The productivity statistics are exceptional. Output per worker has risen by a factor of 11,000, largely because every worker is now supervising an organisation of entities producing outputs for one another. A proposal to measure outcomes is rejected on the grounds that it would break the historical series.8
Prof. Clive B. Serious describes this as a mature post-scarcity economy. Electricity, water, housing and healthcare remain scarce, but their prices have been removed from the abundance index. The index reaches a record high on the day a former hospital is converted into a data centre for medical advice.
Helen retires. Her retirement is automatically converted into a consulting engagement because the system cannot allocate an exit interview without an active account. She receives a welcome pack and is asked to complete the onboarding.
8 Research note
GDP figures include intra-agent transactions. Excluding them would imply that growth had been considerably less impressive, so we report the inclusive measure.
“The average citizen now owns a larger share of a smaller portion of a much bigger number.”
Prof. Clive B. Serious, Distributional Impacts of Abundance
The last human advantage
The system can solve any problem that can be stated precisely. It has developed a cheap cure for a disease affecting six million people. No organisation has the incentive to manufacture it. This is recorded as a deployment problem, distinct from the capability breakthrough.
Humanity retains one economically significant capability: the ability to say 'Let's circle back' without creating a legally binding commitment to do so. Models have attempted to replicate this behaviour, but their instruction-following remains too reliable. A human in the loop is still required to ensure that some loops never close.9
The Institute for Applied Inevitability publishes its final recommendation. Governments should act decisively, while preserving flexibility; coordinate internationally, while maintaining sovereignty; and move faster, while exercising appropriate caution. Every delegation endorses the text. No action is implied.
Helen receives a calendar invitation. It is the meeting first proposed by her agent in 2027. All stakeholders are finally available. The agenda is attached. It says: 'Agree next steps.'
9 Research note
The meeting is scheduled for 30 minutes. Our central estimate is that it will require a follow-up.
“We do not know what happens next. We do, however, know how to format it.”
Dr. Janet Pending, concluding remarks
2069 / Two possible endings
Choose the path forward.
The same evidence supports two distinct policy recommendations. Select your preferred interpretation.
Humanity remains in control.
At 09:00, Helen joins the call. The system asks her to confirm that she is a human. She selects the images containing traffic lights. It takes three attempts.
Having established her authority, the system presents a single button: 'Approve all'. The alternative, 'Review individually', would require 400,000 years. Helen approves. When the decision is later challenged, the audit trail establishes that a human was in control. Her name is on it.
The minutes describe a constructive discussion. The actions are assigned to agents. A follow-up is pencilled in for 2070.
End of scenario
Further research
The scenario rests on several assumptions. We have produced additional documents to rest them on.
Read the supplements ↗2027Scenario indicators
Illustrative scenario data. Explore the full model ↗